CN-121981185-A - Neural network model parameter optimization method based on big data

Abstract

The invention relates to the technical field of edge artificial intelligence and automatic control, in particular to a neural network model parameter optimization method based on big data, which comprises the following steps of firstly, acquiring hardware physical parameters of target execution equipment, real-time physical environment data of a mobile carrier and a service training data set; the method comprises a first step of calculating theoretical reasoning time based on nominal calculation power, a second step of constructing a thermodynamic digital twin model of target execution equipment, a third step of taking the theoretical reasoning time as real reasoning delay, a fourth step of combining the motion speed, the acceleration and the current position of a carrier according to each network configuration, outputting a physical failure risk penalty item according to whether a predicted space position breaks through a safety envelope, a fifth step of constructing accuracy and a thermodynamic loss item based on prediction accuracy and internal temperature, and evaluating and selecting target configuration according to a combination loss function corresponding to a plurality of calculated parameter configuration and issuing the target configuration.

Inventors

• HE YIFAN
• WANG ZHIHUA
• WANG ZONGYUE

Assignees

• 集美大学

Dates

Publication Date

20260505

Application Date

20260403

Claims (7)

1. 1. The neural network model parameter optimization method based on big data is characterized by comprising the following steps of: the method comprises the steps of firstly, acquiring hardware physical parameters of target execution equipment, real-time physical environment data of a mobile carrier and a service training data set, wherein the hardware physical parameters comprise a power consumption heating curve, radiator thermal resistance, a frequency reduction temperature threshold value and a nominal calculation force, and the environment data comprise the current temperature of the equipment, the carrier movement speed, carrier movement acceleration, the current position of the carrier and a safety space envelope curve; training the super-network model based on the service training data set to generate a plurality of network parameter configurations, evaluating the prediction accuracy and the theoretical floating point operand of the network parameter configurations, and calculating the theoretical reasoning time based on the nominal calculation power; Thirdly, constructing a thermodynamic digital twin model of the target execution device, respectively inputting theoretical floating point operand and predicting the internal temperature of the device after continuously running a preset time window by combining the current temperature of the device according to each network parameter configuration in the plurality of network parameter configurations; step four, aiming at each network parameter configuration, respectively combining the carrier movement speed, the carrier movement acceleration and the carrier current position, calculating a predicted space position under the real reasoning delay time, and outputting a physical failure risk penalty item according to whether the predicted space position breaks through a safety space envelope; And fifthly, constructing an accuracy loss item and a thermal loss item based on the prediction accuracy and the equipment internal temperature, constructing a joint loss function by combining the physical failure risk penalty item, evaluating and selecting a target network parameter configuration according to the calculated joint loss function corresponding to the network parameter configurations, and transmitting the target network parameter configuration to target execution equipment.
2. 2. The method for optimizing parameters of a neural network model based on big data according to claim 1, wherein the network parameter configuration comprises a full-scale parameter configuration, a first proportional pruning parameter configuration and a second proportional pruning parameter configuration, and the super-network model outputs a sub-network structure corresponding to the network parameter configuration through a dynamic path selection mechanism.
3. 3. The method for optimizing parameters of a neural network model based on big data according to claim 1, wherein the thermodynamic digital twin model adopts a thermal resistance-heat capacity network architecture, and the thermal resistance-heat capacity network architecture establishes a differential equation of heating power and temperature rise through the power consumption heating curve and the radiator thermal resistance.
4. 4. The neural network model parameter optimization method based on big data according to claim 1, wherein the calculating the true inference delay time based on the down-conversion mechanism comprises: Acquiring the actual working frequency of the target execution equipment at the internal temperature of the equipment, and calculating the actual floating point calculation force at the actual working frequency; dividing the theoretical floating point operand by the actual floating point operand to obtain the real reasoning delay time.
5. 5. The method for optimizing parameters of a neural network model based on big data according to claim 1, wherein the calculating of the predicted spatial position under the real reasoning delay time is characterized by calculating a physical displacement by using a kinematic formula, wherein the physical displacement is equal to a product of the carrier motion speed and the real reasoning delay time, a product of one half of the carrier motion acceleration and the real reasoning delay time square is added, and the physical displacement is added to the current position of the carrier to obtain the predicted spatial position.
6. 6. The neural network model parameter optimization method based on big data according to claim 1, wherein the joint loss function is formed by respectively endowing the accuracy loss term, the thermal loss term and the physical failure risk penalty term with a first weight coefficient, a second weight coefficient and a third weight coefficient to perform linear superposition; The first weight coefficient, the second weight coefficient and the third weight coefficient are subjected to self-adaptive dynamic adjustment based on the internal temperature of the equipment and the carrier movement speed, when the internal temperature of the equipment is higher than or equal to the frequency-reduction temperature threshold value, the second weight coefficient and the third weight coefficient are increased, the first weight coefficient is reduced, and when the internal temperature of the equipment is lower than the frequency-reduction temperature threshold value, the first weight coefficient, the second weight coefficient and the third weight coefficient are kept to be preset initial values.
7. 7. The method for optimizing parameters of a neural network model based on big data according to claim 1, wherein the target execution device is deployed in an autopilot control system or an industrial robot control system, and the safety space envelope is a minimum safety distance constraint for avoiding physical collision.

Description

Neural network model parameter optimization method based on big data Technical Field The invention relates to the technical field of edge artificial intelligence and automatic control, in particular to a neural network model parameter optimization method based on big data. Background With the development of artificial intelligence technology, the deployment of neural network models on edge computing devices is becoming increasingly popular, however, edge devices often face strict heat dissipation restrictions; The parameter optimization of the existing network model often excessively pursues the static big data prediction accuracy, but neglects the heating problem of the equipment caused by continuous high-load operation under the complex working condition; when the internal temperature of the edge equipment reaches a hardware down-conversion temperature threshold value, the system triggers a hardware overheat protection mechanism to reduce the working frequency, which leads to sudden drop of a model push-out management force and causes uncontrollable reasoning delay; therefore, how to provide a neural network model parameter optimization method capable of adapting to the real-time thermodynamic state of the edge device, ensuring the prediction accuracy and avoiding the risk of physical execution delay is a problem to be solved by the skilled person. Disclosure of Invention The invention aims to provide a neural network model parameter optimization method based on big data, which solves the following technical problems: The problem of safety failure risk caused by physical execution delay due to excessive pursuit of static big data prediction accuracy is solved, so that the model can meet the reasoning timeliness requirement according to the real-time thermal state of the edge equipment while ensuring the prediction accuracy, and systematic downtime or disastrous delay caused by chip overheat frequency reduction is effectively avoided, thereby improving the safety and continuous usability under extreme working conditions. The aim of the invention can be achieved by the following technical scheme: the neural network model parameter optimization method based on big data comprises the following steps: the method comprises the steps of firstly, acquiring hardware physical parameters of target execution equipment, real-time physical environment data of a mobile carrier and a service training data set, wherein the hardware physical parameters comprise a power consumption heating curve, radiator thermal resistance, a frequency reduction temperature threshold value and a nominal calculation force, and the environment data comprise the current temperature of the equipment, the carrier movement speed, carrier movement acceleration, the current position of the carrier and a safety space envelope curve; training the super-network model based on the service training data set to generate a plurality of network parameter configurations, evaluating the prediction accuracy and the theoretical floating point operand of the network parameter configurations, and calculating the theoretical reasoning time based on the nominal calculation power; Thirdly, constructing a thermodynamic digital twin model of the target execution device, respectively inputting theoretical floating point operand and predicting the internal temperature of the device after continuously running a preset time window by combining the current temperature of the device according to each network parameter configuration in the plurality of network parameter configurations; step four, aiming at each network parameter configuration, respectively combining the carrier movement speed, the carrier movement acceleration and the carrier current position, calculating a predicted space position under the real reasoning delay time, and outputting a physical failure risk penalty item according to whether the predicted space position breaks through a safety space envelope; And fifthly, constructing an accuracy loss item and a thermal loss item based on the prediction accuracy and the equipment internal temperature, constructing a joint loss function by combining the physical failure risk penalty item, evaluating and selecting a target network parameter configuration according to the calculated joint loss function corresponding to the network parameter configurations, and transmitting the target network parameter configuration to target execution equipment. Preferably, the network parameter configuration includes a total parameter configuration, a first proportion pruning parameter configuration and a second proportion pruning parameter configuration, and the super-network model outputs a sub-network structure corresponding to the network parameter configuration through a dynamic path selection mechanism. Preferably, the thermodynamic digital twin model adopts a thermal resistance-heat capacity network architecture, and the thermal resistance-heat capacity network architec